Say what a create run's failure actually was, per the first consumer

[wormeyman]

The first probe a consumer wrote against this tool found a real defect in it, and
cost four lessons the three documents exist to carry. This is both.

The defect

A create run that wrote its dump under a different name reported:

no dump was written. The most common cause is a factorio_version mismatch,
which makes Factorio skip the mod in silence.

That names a cause the run had already ruled out. The probe raised DUMPED-OK,
and a mod skipped over a version mismatch never runs, so it cannot raise the
sentinel. The 270 KB it had just written was listed in the report's own files
array the whole time. The message sent a reader to info.json for a mod that
demonstrably loaded.

So the message now splits on sentinel_seen, which RunFacts already carried
and the Create arm ignored:

• No sentinel. Nothing ran, so the silent skip really is the first thing to
  check, and the message still says so.
• Sentinel seen. The message says the mismatch is excluded and points at the
  filename instead.

Both arms now name oracle-dump.json rather than saying "no dump". That name is
this tool's contract rather than the game's, and it was documented nowhere -
not in the README, not in docs/, only in a constant in run.rs.

Both new tests were mutation-checked: collapsing the two arms back to the one old
message fails both. One existing test in run.rs asserted contains("no dump")
and now asserts the two properties that distinguish the arms.

The docs

• docs/gotchas.md - the filename contract, and that basis_noise's
  gradient table is a constant of the engine rather than of a version. It
  recovers byte-identically from 2.0.77 and 2.1.14, and scores 512 of 512 exact
  against a fixture first captured on 2.1.11 and re-verified on 2.1.12.
• docs/method.md - a wrong inversion returns a full set of plausible
  numbers. The one-corner form the consumer's own notes describe recovers 2 of
  256 slots, raises nothing, and returns 512 values of plausible size. Plus
  scoring against what the method never read, pricing a limit the method cannot
  see, and a concrete case of a control that was malformed rather than failing.
• README.md - the output filename, where a probe author will look for it.
• CLAUDE.md - the test count was 247 and is 255, and the measured facts
  now record what that probe settled.

Verification

cargo fmt --all -- --check && cargo clippy --all-targets -- -D warnings && cargo test --all-targets

Clean. 255 unit tests and 13 integration tests. All four gates were satisfied
on the machine that ran them, including FACTORIO_ORACLE_PROVENANCE_DIR and
FACTORIO_ORACLE_OLD_FACTORIO, so none of the 13 skipped.

🤖 Generated with Claude Code

https://claude.ai/code/session_016Qg3gYjaG2uEtvv5U6AjY6